GPT Image 2.6 prompt patterns that work with fixed styles
Most prompt advice on the internet was written for a model that had to be told everything: the lighting, the lens, the mood, the render engine. That advice does not transfer to GPT Image 2.6 running behind a fixed style template, and following it is the main reason a generation comes back looking muddy.
Here is the distinction that matters. A style template is a complete visual recipe — lighting direction, material response, focal length, colour grade, framing. It is already committed. Your prompt is not describing an image. It is describing who is in the image, and the template renders that subject in its own look. Once you accept that split, prompt writing collapses from four sentences to one, and the hit rate goes up.
The patterns below come from building and running the 25 templates on WinkImg, all of which render through GPT Image 2.6 Flare. They are about behaviour, not about benchmark scores, because the useful question is "will the next generation be better than this one", not "what is the FID".
Six patterns that work
1. Three nouns beat three paragraphs
The templates that produce the most consistent results are the ones where users type a short list: hair, outfit, one prop. A figurine prompt like "silver bob, oversized cream sweater, holding a tiny game controller" gives the model three concrete anchors. It does not need to invent anything, so it does not invent anything wrong.
Long descriptions fail differently, not better. Every extra clause is another chance for the model to resolve a conflict — a "neon-lit alley" clause inside a template that already sets its own lighting will fight the template and produce something in between.
2. Concrete nouns, not adjectives
"Retro" is a mood word. "Brass goggles" is an object. GPT Image 2.6 renders the object; it interprets the mood, and interpretation is where drift comes from.
Swap abstract words for physical ones:
| Instead of | Write |
|---|---|
| a cool outfit | a black leather jacket with a silver zipper |
| retro vibe | brass goggles and a torn map |
| cyberpunk girl | shaved side, neon-green eye implant, clear raincoat |
| anime style | (delete it — the template is the anime style) |
3. Anchor the subject first, put the action last
Order matters more than people expect. Start with who the subject is, then what they are doing, then where. "A courier in an orange jacket, mid-stride, one hand holding a parcel" reads cleanly. "Mid-stride while holding a parcel, with a courier in orange" makes the model resolve the action before it knows who is acting, and the body proportions suffer.
4. Name the material, not the colour
Colour words get you hue. Material words get you light response, and light response is most of what makes a template look expensive. "Matte black" and "polished chrome" are both dark, and they look nothing alike. In the desk figurine template, adding "vinyl with a soft gloss" changes the surface more than adding any adjective about quality.
5. For GIFs, describe a loop, not a scene
GIF output on 2.6 renders 2 to 16 frames and packs them into a single sheet before encoding, so the loop is consistent frame to frame — but only if the action can return to where it started. "Hair lifting in the wind" loops. "Walking forward across the frame" does not: frame 16 has left the frame 1 position and the loop visibly snaps.
Write loops as oscillations: breathing, blinking, a flag fluttering, a cape swaying, a light pulsing. Anything with a cycle.
6. For a series, change exactly one variable
When you want eight images of the same character in different styles, the style change is the variable. Keep hair, outfit and prop byte-identical between runs and change only the template. The moment you also rewrite the subject line, you lose the ability to tell whether a difference came from the template or the prompt.
This is the same discipline as a controlled test, and it is the only reliable way to build a consistent character series with a single-model tool.
Four patterns that waste a generation
1. Stacking style words
"Cyberpunk steampunk ghibli watercolour" gives you a blur. Inside a template, the style is already fixed by the recipe; adding style words does not add style, it adds a conflicting instruction. Pick one template and let it do its job.
2. Camera parameters
"Cinematic, 85mm, f/1.4, shot on Portra 400" is noise here. The template owns focal length and depth of field. Camera language was essential when the model had no idea what an image should look like; it is decoration now.
3. Empty quality tokens
"4k, ultra detailed, masterpiece, best quality, trending on artstation" are tokens trained to correlate with nothing in particular. They cost you nothing in output quality and cost you a lot in clarity.
4. Negations
"no extra fingers, no text, no watermark" is a weak instruction in image models generally. Restate it positively instead: "both hands visible, resting on the table". Positive framing gives the model something to render; negation gives it something to avoid, which is not the same operation.
What actually changed with 2.6
Two things, and they are both boring in a useful way.
Speed changed the iteration economics. Flare returns a typical frame in about six seconds. Six seconds is short enough that you try a second phrasing instead of accepting a mediocre result, and that behaviour — trying again — does more for output quality than any prompt trick.
GIF changed what a template can output. A style used to end as a still. Now a style can be a short loop, which means the composition has to be designed for motion rather than cropped for a still. That is a different skill, and pattern 5 above is most of it.
The consistency story is more mixed than the marketing suggests. Flare is the fast variant; the consistency-tuned variant is Sunburst, and it is not exposed as an option on WinkImg. Within one template, material and lighting stay stable because the template is fixed — that part is genuinely reliable. Subject drift across many rounds of editing is reduced compared with earlier models, but it is not gone, which is why pattern 6 exists.
The short version
Fixed template, one line, three concrete nouns, one variable changed at a time. If a generation misses, change a noun — do not add a sentence. That single rule accounts for most of the difference between people who get good results out of GPT Image 2.6 on the first try and people who keep regenerating the same paragraph.
If you want to see how the templates behave before writing anything, the style library has a page per style with the exact look each one produces. Pick the style first. The prompt is the easy part.